Session Item

The quantitative comparison of MRI intensities between sessions, patients, and machines requires a normalization. There is currently no consensus about the optimal normalization approach. Here we compared the consistency of normalized intensity values within different tissue types in brain MRI for three commonly used intensity normalization techniques and a newly developed method.

We analyzed the publicly available dataset CC-359, containing 359 T1-w brain MRI of different healthy individuals from 6 different scanners. Preprocessing of the images consisted of brain extraction and N4 bias field correction.

Using the FSL FAST algorithm, all images were automatically segmented into three tissue types, forming the three regions of interest (ROI): cerebrospinal fluid (CSF), gray matter (GM) and white matter (WM). Voxels were assigned to a region if the algorithm reported 100% certainty.

The evaluated normalization techniques were: piecewise linear normalization known as Nyul, z-score transformation on all brain intensities, normalization based on the WM intensity peak known as WhiteStripe, and a custom normalization method.

The custom method isolates the homogeneous parts of the image by excluding voxels whose local surroundings have a high change in intensity. In the remaining image, the two peak intensities are detected and interpreted as CSF and WM. Then, the intensities of the entire image are linearly transformed, mapping the CSF peak intensity to 0 and the WM peak intensity to 100.

For each normalization method and ROI, the Jensen-Shannon distance (JSD) was calculated between each subject’s histogram and the average histogram among all subjects. A lower JSD across subjects correspond to less variability between subjects and thus a more consistent normalization of the intensities in the ROI.

For each normalization method and each ROI, the histograms of each subject, along with the average histogram of all subjects, are shown in figure 1. All normalization methods resulted in more consistent intensity values than no normalization (figure 2). Nyul normalization achieved the lowest median JSD for each ROI. Among the other techniques, the lowest median JSD for WM was achieved in WhiteStripe and for CSF and GM in z-score. For each ROI, all pairwise comparisons of JSD values between normalizations were statistically significant (p<0.05; Wilcoxon), except for CSF between the custom method and z-score (p=0.83).

Nyul normalization performed best in each ROI. However, it has caveats not addressed in this evaluation, e.g. it affects the image texture. The normalization methods which are based on known tissue characteristics (WhiteStripe and custom) performed well in at most 2 out of 3 ROI. The same was observed for z-score. This suggests that a nonlinear normalization is necessary to achieve consistent intensities for all brain tissue types. This knowledge will be used to further develop the methods.